Optimization of O2O orders delivery path based on deep reinforcement learning

During the COVID-19 pandemic, the Online to Offline (O2O) business sector plays a crucial role in safeguarding people's livelihoods. Millions of transactions occur daily on online logistics ordering platforms. The allocation of orders and the path planning of delivery vehicles present significant challenges to the O2O platforms. This paper presents a collaborative optimization deep reinforcement learning framework (CODRL) based on an encoder-decoder architecture. A node pairing embedding is designed in the encoder to capture the structural dependency relationship between merchant nodes and customer nodes in the same order. Subsequently, the vehicle decoder is utilized to select a delivery vehicle, and the node decoder is used to specify a node that requires service by the delivery vehicle. The preliminary experimental results, utilizing randomly generated order data, demonstrate that the proposed CODRL outperforms all other baselines by at least 5.67% across all settings. Furthermore, the performance enhancement of the node pairing embedding, designed in the encoder, is evaluated through ablation study.