Multi-Agent Resource Balancing With Heterogeneous Graph Neural Networks

Marine transportation plays an indispensable role in international trade,most of the marine commodities are transported by container.However,the development of international trade is extremely unbalanced.The global economic costs on container management and repositioning is very high.The empty container repositioning problem becomes one of the key issues in marine transportation.Developing reasonable strategies for empty container repositioning can not only reduce the costs of cargo loading and transportation,and reduce the shortage of container resources,but also improve the efficiency of marine logistics transportation and reduce port congestion,so as to indirectly improve resource utilization and promote the development of international trade.To solve the problems of simple source of information and the lack of cooperation mechanism in existing methods,this paper presents a multi-agent cooperative resource balancing model based on heterogeneous graph neural networks.Firstly,this model considers more ocean shipping information by building a heterogeneous ocean transportation network;secondly,a multi-agent cooperative mechanism is proposed by combining graph neural network with multi-agent reinforcement learning.Specifically,the contributions of the method proposed in this paper are listed as follow:(1)We propose a new heterogeneous ocean transportation network for empty container repositioning.To extract information from complex ocean transportation network more efficiently,we construct a heterogeneous ocean transportation network by taking ports and vessels as the nodes,the accessibility within a single route as the edges according to the relationship among the ports,the vessels and the routes.Compared with the original homogeneous ocean transportation network,this proposed heterogeneous network can capture more shipping information.(2)We propose a new multi-agent cooperation model with heterogeneous graph neural networks,which is called heterogeneous graph collaboration model.In this model,graph neural network is applied to automatically learn the node's representation,and model the relationship(influence)between agents and their neighborhoods through graph attention mechanism.Compared with the existing model,this model can automatically learn the interaction of neighboring agents,which could facilitate agent's cooperation.To verify this model,we build a multi-agent simulation system for empty container repositioning.Extensive experiments are conducted within this system,the experimental results show that our model outperform the state-of-the-art methods.Compared with the benchmark model,the order fulfillment rate of the model is increased by 15%,and the costs of empty container repositioning is reduced by 18%.