A Multi-Agent Deep Reinforcement Learning Method with Diversified Policies for Continuous Location of Express Delivery Stations Under Heterogeneous Scenarios

Rational location planning of express delivery stations (EDS) is crucial for enhancing the quality and efficiency of urban logistics. The spatial heterogeneity of logistics demand across urban areas highlights the importance of adopting a scientific approach to EDS location planning. To tackle the issue of strategy misalignment caused by heterogeneous demand scenarios, this study proposes a continuous location method for EDS based on multi-agent deep reinforcement learning. The method formulates the location problem as a continuous maximum coverage model and trains multiple agents with diverse policies to enable adaptive decision-making in complex urban environments. A direction-controlled continuous movement mechanism is introduced to facilitate an efficient search and high-precision location planning. Additionally, a perception system based on local observation is designed to rapidly capture heterogeneous environmental features, while a local–global reward feedback mechanism is established to balance localized optimization with overall system benefits. Case studies conducted in Fuzhou, Fujian Province and Shenzhen, Guangdong Province, China, demonstrate that the proposed method significantly outperforms traditional heuristic methods and the single-agent deep reinforcement learning method in terms of both coverage rate and computational efficiency, achieving an increase in population coverage of 9.63 and 15.99 percentage points, respectively. Furthermore, by analyzing the relationship between the number of stations and coverage effectiveness, this study identifies optimal station configuration thresholds for different urban areas. The findings provide a scientific basis for investment decision-making and location planning in EDS construction.