| 基于深度强化学习求解容量限制车辆路径问题 |
| 点此下载全文 |
| 引用本文：梁晓茹,范静.基于深度强化学习求解容量限制车辆路径问题[J].上海第二工业大学(中文版),2026,43(1):81-88 |
| 摘要点击次数: 311 |
| 全文下载次数: 18 |
|
|
| 中文摘要:本文提出一种融合循环神经网络 (recurrent neural network, RNN) 时序动态建模能力、注意力机制、动态特征筛选机制及深度 Q 网络 (deep Q-network, DQN) 强化学习框架的 RAD 算法, 用于求解有容量限制的车辆路径问题(constrained vehicle routing problem, CVRP)。该算法中, RNN 的门控循环单元 (gated recurrent unit, GRU) 用于捕捉车辆载重状态与客户需求的时序依赖关系, 注意力机制自适应聚焦可服务客户节点, DQN 的双网络架构与约束
惩罚策略则用于优化路径决策。对于随机数据集, RAD 算法相较于 EL-DRL、AKS 与 PAN-CAS, 在总路径长度、平均奖励和计算效率上均有明显优势。在 Solomon 数据集上, RAD 算法的结果验证了其在复杂容量约束条件下的高效求解能力。此外, 消融实验表明, RAD 算法以少量的时间代价即可获得更优的运输方案。 |
| 中文关键词:深度强化学习 深度 Q 网络 车辆路径问题 注意力机制 |
| |
| Deep Reinforcement Learning Algorithm for Solving the Capacitated Vehicle Routing Problem |
|
|
| Abstract:The RAD algorithm that integrates the temporal dynamic modeling capability, attention mechanism, dynamic feature selection mechanism, and deep Q-network (DQN) reinforcement learning framework of recurrent neural network (RNN), is proposed for solving the capacity constrained vehicle routing problem (CVRP). The gated recurrent unit (GRU) in RNN captures the temporal dependency relationship between vehicle load status and customer demand, the attention mechanism is used to adaptively focus on service-oriented customer nodes, and the DQN’s dual network architecture and constraint penalty strategy can optimize path decision-making. For the random datasets, the RAD algorithm shows significant advantages over EL-DRL, AKS, and PAN-CAS in terms of total route length, average reward, and computational efficiency. For Solomon datasets, the results of the RAD algorithm validate its efficient solving capability under complex capacity constraints. Furthermore, the ablation experiments demonstrate that the RAD algorithm can achieve a superior transportation plan with only a minimal time cost. |
| keywords:deep reinforcement learning deep Q-network vehicle routing problem attention mechanism |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|