PTB: A deep reinforcement learning method for flexible logistics service combination problem with spatial-temporal constraint

The Fourth Party Logistics Service platform aims to provide customers with one-stop logistics services by combining logistics services from different logistics service providers. However, it is an important challenge to quickly build the logistics service combination with the maximum overall profit under the premise of satisfying spatial–temporal constraints. The traditional heuristic algorithm has the problem of low computing efficiency and insufficient adaptability to dynamic environment and demand uncertainty when dealing with the problem of large-scale logistics service combination with spatial–temporal constraints and uncertainty of service. Therefore, we propose the Proximal policy optimization algorithm integrating the Transformer network and the Bayesian networks (PTB). Firstly, a variable length sequence module based on the greedy strategy is used in the environment to generate the status information of the order, to solve the problem that the traditional reinforcement learning environment cannot handle any number of service providers. Then, a policy network integrating the Transformer and Bayes network is used to further optimize the accuracy and reliability of scheduling decisions. Meanwhile the PPO algorithm is used to update the network and constantly optimize the scheduling decisions of agents. Finally, the experimental results on four sets of logistics order scenes show that PTB can effectively solve the logistics service combination problem with spatial–temporal constraints, and the results of most logistics service combination tasks are better than other baseline models, and in large-scale order scenes, it has far more computational efficiency than the traditional heuristic algorithm, and shows good generalization ability.