耦合改进图注意力网络与深度强化学习的公共服务设施智能化选址方法
王 中(2001— ),男,湖南长沙人,硕士生,主要从事空间优化相关研究。E-mail: 51253901082@stu.ecnu.edu.cn。 |
Copy editor: 蒋树芳 , 黄光玉
收稿日期: 2024-01-21
修回日期: 2024-02-18
网络出版日期: 2024-11-07
基金资助
国家重点研发计划项目(2022YFB3903700)
国家重点研发计划项目(2022YFB3903704)
An Intelligent Site Selection Approach for Public Service Facilities Coupled with Improved Graph Attention Network and Deep Reinforcement Learning
Received date: 2024-01-21
Revised date: 2024-02-18
Online published: 2024-11-07
Supported by
National Key Research and Development Program of China(2022YFB3903700)
National Key Research and Development Program of China(2022YFB3903704)
在当今城市化快速发展的背景下,公共服务设施的合理选址对于提升城市居民生活质量和提供高效服务至关重要。然而,现有常用的设施选址方法往往未能满足复杂及大规模的现实场景中对于其性能及效率上需求。为弥补这些不足,本文旨在提出一种耦合设施选址图注意力网络(Facility Location Allocation Graph Attention Network, FLA-GAT)和深度强化学习(Deep Reinforcement Learning, DRL)算法的通用图强化选址模型(Graph-Deep-Reinforcement-Learning Facility Location Allocation Model, GDRL-FLAM),通过图表示和REINFORCE算法协同解决公共服务设施智能化选址问题。为了验证提出模型的性能及效率,研究在随机生成的20、50和100个点的数据集上进行训练,并完成了相应的测试实验,结果表明: ① 在20、50和100个点的测试实例上,GDRL-FLAM模型相较于遗传算法(Genetic Algorithm, GA))性能上提升了11.79%到14.49%;在150和200个点的测试实例上,提升了1.52%到9.35%。并且随着训练集规模的增大,模型在大规模数据集上表现出更强的泛化能力; ② GDRL-FLAM模型可以在简单场景中掌握选址策略,并使其适应到更复杂的场景,展示出了其优异的迁移学习能力;③ 在新加坡的案例研究中,GDRL-FLAM模型相较于GA在性能表现上提升了1.01%到10.75%; ④ 在所有的测试及实验中,GDRL-FLAM模型在效率方面相较于GA都展示出了成倍的提升。总的来说,本研究揭示了GDRL-FLAM模型在公共服务设施选址问题上的潜在应用价值,尤其是其泛化能力及迁移学习能力为未来设施选址问题的高效解决提供了新的思路和方法。此外,该模型经过微调也可以适用于不同的空间优化问题中。最后,本研究探讨了该模型的不足之处以及下一步的研究方向。
王中 , 曹凯 . 耦合改进图注意力网络与深度强化学习的公共服务设施智能化选址方法[J]. 地球信息科学学报, 2024 , 26(11) : 2452 -2464 . DOI: 10.12082/dqxxkx.2024.240044
In the context of the rapid development of urbanization, the reasonable selection of locations for public service facilities is critical for delivering efficient services and enhancing the quality of urban residents' lives. However, prevailing approaches for allocation of public service facilities often fall short of meeting the demands on their performance and efficiency in complex and large-scale real-world scenarios. To address these issues, this article proposed a novel Graph-Deep-Reinforcement-Learning Facility Location Allocation Model (GDRL-FLAM), coupling a Facility Location Allocation Graph Attention Network (FLA-GAT) with a Deep Reinforcement Learning (DRL) algorithm. This proposed model tackled the location allocation problem for public service facilities based on graph representation and the REINFORCE algorithm. To assess the performance and efficiency of the proposed model, this study conducted experiments based on randomly generated datasets with 20, 50, and 100 points. The experimental results indicated that: (1) For the tests with 20, 50, and 100 points, the GDRL-FLAM model exhibited a significant improvement ranging from 11.79% to 14.49% compared to the Genetic Algorithm (GA) which is one of the commonly used heuristic algorithms for addressing location allocation problems. For the tests with 150 and 200 points, the improvement ranged from 1.52% to 9.35%. Moreover, with the increase in the size of the training set, the model also demonstrated enhanced generalizability on large-scale datasets; (2) The GDRL-FLAM model showed strong transfer learning ability to obtain the location allocation strategies in simple scenarios and adapt them to more complex scenarios; (3) In the case study of Singapore, the GDRL-FLAM model outperformed GA significantly, achieving obvious improvements ranging from 1.01% to 10.75%; (4) In all these abovementioned tests and experiments, the GDRL-FLAM model showed substantial improvement in efficiency compared to GA. In short, this study demonstrated the potential of the proposed GDRL-FLAM model in addressing the location allocation issues for public service facilities, due to its generalization and transfer learning abilities. The proposed GDRL-FLAM could also be adapted to solve other spatial optimization problems. Finally, the article discussed the limitations of the model and outlined potential directions for future research.
表1 参数设置Tab. 1 Parameter setting |
参数 | 值 | 参数 | 值 |
---|---|---|---|
Epoch | 200 | 学习率衰减系数 α | 0.98 |
编码层数L | 4 | ε | 0.90 |
学习率 | 0.000 1 | ε 衰减系数 β | 0.99 |
节点嵌入维度hx | 128 | 边嵌入维度he | 64 |
表2 不同节点数量下的GDRL-FLAM性能评估Tab. 2 Performance evaluation of the GDRL-FLAM with different numbers of nodes |
方法 | FLP20 | FLP50 | FLP100 | ||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
目标值 | DI/% | 时间 | 目标值 | DI/% | 时间 | 目标值 | DI/% | 时间 | |||
GA | 4.411 | 0.00 | 3 h | 8.700 | 0.00 | 5 h | 14.081 | 0.00 | 12 h | ||
GDRL-FLAM (ε-greedy) | 3.773 | 14.33 | 11 min | 7.593 | 12.56 | 27 min | 12.400 | 11.79 | 58 min | ||
GDRL-FLAM (Sample) | 3.766 | 14.49 | 24 min | 7.566 | 12.87 | 53 min | 12.373 | 11.99 | 2 h |
表3 FLP150泛化性能测试结果Tab. 3 Generalization performance test results for FLP150 |
方法 | GDRL-FLAM50 | GDRL-FLAM100 | |||||
---|---|---|---|---|---|---|---|
目标值 | DI/% | 时间/h | 目标值 | DI/% | 时间/h | ||
GA | 17.787 | 0.00 | 11 | 17.787 | 0.00 | 11 | |
GDRL-FLAM (ε-greedy) | 17.406 | 1.97 | 2 | 16.273 | 8.34 | 2 | |
GDRL-FLAM (Sample) | 16.922 | 4.69 | 5 | 16.094 | 9.35 | 5 |
表4 FLP200泛化性能测试结果Tab. 4 Generalization performance test results for FLP200 |
方法 | GDRL-FLAM50 | GDRL-FLAM100 | |||||
---|---|---|---|---|---|---|---|
目标值 | DI/% | 时间/h | 目标值 | DI/% | 时间/h | ||
GA | 15.282 | 0.00 | 10 | 15.282 | 0.00 | 10 | |
GDRL-FLAM (ε-greedy) | 15.026 | 1.52 | 1 | 14.613 | 4.23 | 1 | |
GDRL-FLAM (Sample) | 14.792 | 3.06 | 4 | 13.835 | 9.33 | 4 |
表5 迁移学习测试结果Tab. 5 The test results of transfer learning |
方法 | 预训练 | 从头开始 | |||||
---|---|---|---|---|---|---|---|
目标值 | DI/% | 时间 | 目标值 | DI/% | 时间 | ||
GA | 8.700 | 0.00 | 5 h | 8.700 | 0.00 | 5 h | |
GDRL-FLAM (ε-greedy) | 7.588 | 12.63 | 26 min | 7.593 | 12.56 | 27 min | |
GDRL-FLAM (Sample) | 7.561 | 12.94 | 53 min | 7.566 | 12.87 | 53 min |
表6 新加坡案例实验结果Tab. 6 The results of case study in Singapore |
方法 | GDRL-FLAM50 | GDRL-FLAM100 | |||
---|---|---|---|---|---|
目标值 | DI/% | 目标值 | DI/% | ||
GA | 16.715 | 0.00 | 16.715 | 0.00 | |
GDRL-FLAM (ε-greedy) | 16.546 | 1.01 | 15.590 | 6.73 | |
GDRL-FLAM (Sample) | 15.900 | 4.88 | 15.031 | 10.75 |
[1] |
高军波, 周春山. 西方国家城市公共服务设施供给理论及研究进展[J]. 世界地理研究, 2009, 18(4):81-90.
[
|
[2] |
国家发展改革委. 关于印发《“十四五”公共服务规划》的通知[EB/OL].(2022-01-10)[2024-01-21]. https://www.gov.cn/zhengce/zhengceku/2022-01/10/content_5667482.html.
[National Development and Reform Commission. Notice on Issuing the "14th Five-Year Plan" Public Service Plan[EB/OL].(2022-01-10)[2024-01-21]. https://www.gov.cn/zhengce/zhengceku/2022-01/10/content_5667482.html.]
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
关亚宾, 马瑞, 孔云峰. 城市社区便民服务中心选址模型研究[J]. 地球信息科学学报, 2023, 25:2164-2177.
[
|
[13] |
翟石艳, 孔云峰, 宋根鑫, 等. 面向15 min生活圈的城市服务设施规划模型与实验[J]. 地理学报, 2023, 78(6):1484-1497.
[
|
[14] |
|
[15] |
|
[16] |
席裕庚, 柴天佑, 恽为民. 遗传算法综述[J]. 控制理论与应用, 1996, 13(6):697-708.
[
|
[17] |
黎夏. 协同空间模拟与优化及其在快速城市化地区的应用[J]. 地球信息科学学报, 2013, 15(3):321-327.
[
|
[18] |
李凯文, 张涛, 王锐, 等. 基于深度强化学习的组合优化研究进展[J]. 自动化学报, 2021, 47(11):2521-2537.
[
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
[34] |
王玉璟, 孔云峰. 义务教育就近入学优化建模研究[J]. 地球信息科学学报, 2021, 23(9):1608-1616.
[
|
[35] |
|
[36] |
|
[37] |
|
[38] |
|
[39] |
|
[40] |
|
[41] |
|
[42] |
|
[43] |
刘萌伟, 黎夏. 基于Pareto多目标遗传算法的公共服务设施优化选址研究——以深圳市医院选址为例[J]. 热带地理, 2010, 30(6):650-655.
[
|
[44] |
|
[45] |
|
/
〈 |
|
〉 |