Improving the state-of-the-art in the Traveling Salesman Problem: An Anytime Automatic Algorithm Selection

This work presents a new metaheuristic for the euclidean Traveling Salesman Problem (TSP) based on an Anytime Automatic Algorithm Selection model using a portfolio of five state-of-the-art solvers. We introduce a new spatial representation of nodes, in the form of a matrix grid, avoiding costly calc...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Expert systems with applications Jg. 187; S. 115948
Hauptverfasser: Huerta, Isaías I., Neira, Daniel A., Ortega, Daniel A., Varas, Vicente, Godoy, Julio, Asín-Achá, Roberto
Format: Journal Article
Sprache:Englisch
Veröffentlicht: New York Elsevier Ltd 01.01.2022
Elsevier BV
Schlagworte:
Algorithm selection problem
Algorithms
Anytime behavior approach
Artificial neural networks
Combinatorial optimization
Datasets
Heuristic methods
Machine learning
Ranking
Representations
Solvers
Traveling salesman problem
TSP
TSP
Anytime behavior approach
Combinatorial optimization
Algorithm selection problem
ISSN:0957-4174, 1873-6793
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:This work presents a new metaheuristic for the euclidean Traveling Salesman Problem (TSP) based on an Anytime Automatic Algorithm Selection model using a portfolio of five state-of-the-art solvers. We introduce a new spatial representation of nodes, in the form of a matrix grid, avoiding costly calculation of features. Furthermore, we use a new compact staggered representation for the ranking of algorithms at each time step. Then, we feed inputs (matrix grid) and outputs (staggered representation) into a classifying convolutional neural network to predict the ranking of the solvers at a given time. We use the available datasets for TSP and generate new instances to augment their number, reaching 6,689 instances, distributed into training and test sets. Results show that the time required to predict the best solver is drastically reduced in comparison to previous traditional feature selection and machine learning methods. Furthermore, the prediction can be obtained at any time and, on average, the metasolver is better than running all the solvers separately on all the datasets, obtaining 79.8% accuracy. •A new metaheuristic for the TSP based on Automatic Algorithm Selection is proposed.•We consider the anytime nature of the solution computation by the solvers in our work.•A spatial representation of instances is presented, avoiding calculation of features.•Our approach outperforms state-of-the-art solvers in public data sets.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0957-4174
1873-6793
DOI:10.1016/j.eswa.2021.115948