2025年03期 v.46 43-51页
黄烁 张学习 谢兴旺 张涛
（广东工业大学，广东 广州 510006）
摘要：针对传统注意力网络模型在选址路径问题中无法有效保留图结构信息的问题，提出一种强化图注意力网络模型。首先，在保留图结构信息的前提下，编码器通过注意力机制提取图结构中的节点信息，以获得节点的高维特征表示及整体图特征信息；然后，解码器利用门控循环单元有效捕获节点序列中的时间依赖关系，并通过逐步解码的方式获取完整解；最后，引入额外的价值网络评估每个动作的价值，以引导策略更新，提升训练效率。实验结果表明，该强化图注意力网络模型能够快速获取选址路径问题的高质量解。
关键词：选址路径问题；强化学习；图注意力网络；价值网络；图结构
中图分类号：TP391 文献标志码：A 文章编号：1674-2605(2025)03-0007-09
DOI：10.12475/aie.20250307 开放获取
Application of Reinforced Graph Attention Network Model for Location Routing Problem
HUANG Shuo ZHANG Xuexi XIE Xingwang ZHANG Tao
(Guangdong University of Technology, Guangzhou 510006, China)
Abstract: To address the limitation of traditional attention network models in effectively preserving graph structural information for Location Routing Problems, this paper proposes a reinforced graph attention network model. First, while retaining graph structural information, the encoder extracts node information from the graph structure through attention mechanisms to obtain high-dimensional feature representations of nodes and global graph feature information. Then, the decoder utilizes Gated Recurrent Units to effectively capture temporal dependencies in node sequences and acquires complete solutions via step-by-step decoding. Finally, an auxiliary Value Network is introduced to evaluate the value of each action, guiding policy updates to enhance training efficiency. Experimental results demonstrate that this reinforced graph attention network model can rapidly obtain high-quality solutions for LRP.
Keywords: location routing problem; reinforcement learning; graph attention network; value network; graph structure
