摘要
现有基于神经网络的车辆路径求解方法通常假设训练和测试实例服从相同分布(即,均匀分布),从而导致跨分布场景的泛化能力弱。为此,本文提出一种基于领域自适应强化学习的跨分布车辆路径求解方法,核心在于领域自适应。该方法设计一种基于分布引导的自适应策略网络(DGATP),并将其嵌入端到端深度强化学习框架(DRL),以解决跨分布车辆路径求解问题。具体的是,首先,构建分布识别模块,以感知源域与目标域的差异并进行特征提取和分布识别;其次,建立门控融合网络,以自适应地加权与融合不同分布的特征;最后,设计感知注意力解码器,以生成路由策略。基于两个代表性深度模型的实验结果表明,与传统方法相比,DGATP性能在跨分布场景下取得显著提升,展现优异的泛化性和通用性。
Most existing neural network-based vehicle routing methods usually assume identical distributions(i.e.,a uniform distributions)across instances,which makes them difficult to generalize to cross-distribution scenarios.To this end,this paper proposes a cross-distribution vehicle routing approach based on domain-adaptive reinforcement learning,with domain adaptation as its core principle.The proposed method involves designing a Distribution-Guided Adaptive Policy Network(DGATP)and embedding it within an end-to-end deep reinforcement learning(DRL)framework to address cross-distribution vehicle routing problems.Specifically,a distribution recognition module is first constructed to perceive differences between the source and target domains by performing feature extraction and distribution identification.Then,a gated fusion network is established to adaptively weight and fuse features from different distributions.Finally,a perception-aware attention decoder is designed to generate routing policies.Experimental results based on two representative deep learning models demonstrate that the DGATP approach significantly outperforms traditional methods in cross-distribution scenarios,exhibiting excellent generalization and universality.
作者
金友龙
夏大文
JIN Youlong;XIA Dawen(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China)
出处
《智能计算机与应用》
2025年第12期17-22,共6页
Intelligent Computer and Applications
基金
贵州省高等学校大数据分析与智能计算重点实验室(黔教技[2023]012号)。
关键词
车辆路径问题
深度强化学习
DGATP
门控融合网络
领域自适应
vehicle routing problems
deep reinforcement learning
DGATP
gate-controlled fusion network
domain adaptation
