BiGNN: Bipartite graph neural network with attention mechanism for solving multiple traveling salesman problems in urban logistics

•Learn to Branch: A new approach is proposed to optimize the branch strategy in Branch-and-bound.•Deep Learning Model: The new model Bipartite Graph Neural Network with Attention Mechanism is proposed to learn the branch strategies.•Comparative Analysis: The models outperform other machine learning...

Full description

Saved in:
Bibliographic Details
Published in:International journal of applied earth observation and geoinformation Vol. 129; p. 103863
Main Authors: Liang, Haojian, Wang, Shaohua, Li, Huilai, Zhou, Liang, Zhang, Xueyan, Wang, Shaowen
Format: Journal Article
Language:English
Published: Elsevier B.V 01.05.2024
Elsevier
Subjects:
algorithms
Attention mechanism
Bipartite graph
Branch-and-bound
Graph neural network
Multiple traveling salesman problem
spatial data
Attention mechanism
Branch-and-bound
Bipartite graph
Multiple traveling salesman problem
Graph neural network
ISSN:1569-8432, 1872-826X
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•Learn to Branch: A new approach is proposed to optimize the branch strategy in Branch-and-bound.•Deep Learning Model: The new model Bipartite Graph Neural Network with Attention Mechanism is proposed to learn the branch strategies.•Comparative Analysis: The models outperform other machine learning algorithms.•Real-World Application: The method is used to solve the Surat Road network data. The multiple traveling salesman problems (MTSP), which arise from real world problems, are essential in urban logistics. Variations such as MinMax-MTSP and Bounded-MTSP aim to distribute workload evenly among salesmen and impose constraints on visited cities, respectively. Branch-and-bound (B&B) provides an exact algorithm solution for these problems. The Learn to Branch (L2B) approach guides branch node selection through deep learning. In this study, we utilize mathematical modeling of Bipartite Graph Neural Network (BiGNN) and an attention mechanism to support B&B in exploring solution spaces through imitation learning. The problems are framed to formulate mixed integer linear programming, which is different from conventional algorithms. It is proposed that a bipartite graph network approach makes a feature representation by setting a structure of constraints and variables. Experimental results showed that our model can generate more accurate solutions than three benchmark models. The BiGNN model can effectively learn the strong branch strategy, which reduces solution time by replacing complex calculations with fast approximations. Additionally, the small-scale instances model can be applied to larger-scale ones.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
ISSN:1569-8432
1872-826X
DOI:10.1016/j.jag.2024.103863