Research And Application Of Terminal Logistics Distribution Based On Deep Reinforcement Learning

With the rapid development of e-commerce and express delivery industry,the end-to-end logistics distribution has become an essential part of the modern logistics system.However,consumers’ demands for logistics services continue to increase,while domestic logistics costs are also rising.Therefore,the rapid development of optimal delivery routes has become an urgent problem that needs to be solved.Traditional path planning algorithms experience an exponential increase in calculation time as the problem size grows,making it difficult to meet the requirements of modern logistics.Based on the development of artificial intelligence technology,path planning models using deep learning have emerged,which can quickly calculate the optimal route.By optimizing the logistics distribution path,it not only helps to ensure timely and accurate delivery for consumers,but also improves the delivery efficiency of logistics enterprises and reduces operating costs,thus promoting the sustained and steady development of the enterprise.According to the characteristics of end-to-end logistics distribution,this study employs an Encoder-Decoder model architecture to design a model.Specifically,the PointerNetworks(PtrNet)is used as the underlying network model,and further improvements are made by introducing techniques such as Multi-Head Attention(MHA),Transformer,and Graph ConvolutionalNetwork(GCN)to enhance the model’s ability to adapt to real-world end-to-end logistics distribution scenarios.Additionally,the model is trained using deep reinforcement learning techniques to enable it to explore the optimal path beyond the constraints of the training samples.To address the insufficient informatization of last-mile logistics,this thesis also designs and develops a last-mile logistics delivery system.The main tasks are as follows:(1)A path planning model called MHA-PtrNet based on deep reinforcement learning is proposed to address the issue of end-to-end logistics distribution.First,the end-to-end logistics distribution problem is regarded as a sequence-to-sequence mapping problem,and the pointer network is used to solve the problem.Based on this,the model training uses the Actor-Critic algorithm in deep reinforcement learning to enable the model to explore the optimal path beyond the constraints of the training samples.Then,the multi-head attention mechanism is employed to embed the data,helping the model understand the relationships between data and improve its performance.Finally,a comparative experiment is conducted using a 2D Gaussian distribution dataset and the TSPLIB dataset based on actual engineering.The results show that the proposed model has the advantages of fast solving speed and good performance on small and medium-scale data.(2)A path planning model called GCN-PtrNet based on graph convolutional networks is proposed to address the issue of poor solution quality in solving large-scale data.First,using graph convolutional networks to encode global data,and input the encoded data as context nodes into the model,enabling the model to utilize the graph structure information in the data.Then,a Transformer encoder-decoder structure is used to directly embed the data to improve the model’s efficiency.In addition,the Rollout algorithm is employed as a benchmark in deep reinforcement learning to increase the model’s training speed.Finally,a comparative experiment is conducted using a 2D Gaussian distribution dataset and the actual road network dataset of Jinan’s city center.The results show that the proposed model can be successfully applied to real end-to-end logistics distribution scenarios and has practical application value.(3)To address the issue of insufficient information technology in end-to-end logistics distribution,we have designed and developed an end-to-end logistics distribution system.The system’s main functions include logistics order management,employee management,vehicle management,and other functions to help users better manage logistics distribution.Additionally,we have integrated the path planning model into the system,combining research with practical applications.With this system,users can conveniently plan logistics distribution routes.