Adaptive recommendation model using meta-learning for population-based algorithms

Xianghua Chu, Fulin Cai, Can Cui, Mengqi Hu, Li Li, Quande Qin

    Research output: Contribution to journalArticle

    Abstract

    To efficiently solve complex optimization problems, numerous population-based meta-heuristics and extensions have been developed. However, the performances of the algorithms vary depending on the problems. In this research, we propose an Adaptive Recommendation Model (ARM) using meta-learning to identify appropriate problem-dependent population-based algorithm. In ARM, the algorithms are adaptively selected by mapping the problem characteristics to the algorithm performance. Since the meta-features extracted and meta-learner adopted would significantly affect the system performance, 18 meta-features including statistical, geometrical and landscape features are extracted to characterize optimization problem spaces. Both instance-based and model-based learners are investigated. Two performance metrics, Spearman's rank correlation coefficient and success rate are used to evaluate the accuracy of optimizer's ranking prediction and the precision of the best optimizer recommendation. The proposed ARM is compared against population-based algorithms with distinct search capabilities such as PSO variants, non-PSO population-based optimizers, hyper-heuristics and ensemble methods. Benchmark functions and real-world problems with various properties are adopted in the experiments. Experimental results reveal the extendibility and effectiveness of ARM on the diverse tested problems in terms of solution accuracy, ranking and success rate.

    LanguageEnglish (US)
    Pages192-210
    Number of pages19
    JournalInformation Sciences
    Volume476
    DOIs
    StatePublished - Feb 1 2019

    Fingerprint

    Meta-learning
    Recommendation System
    Heuristic methods
    Recommender Systems
    Recommender systems
    Heuristic Method
    Global optimization
    Global Optimization
    Particle swarm optimization (PSO)
    Learning algorithms
    Particle Swarm Optimization
    Learning Algorithm
    Recommendations
    Ranking
    Hyper-heuristics
    Optimization Problem
    Spearman's coefficient
    Ensemble Methods
    Model
    Performance Metrics

    Keywords

    • Algorithm selection
    • Global optimization
    • Meta-learning
    • Population-based algorithm
    • Recommendation system

    ASJC Scopus subject areas

    • Software
    • Control and Systems Engineering
    • Theoretical Computer Science
    • Computer Science Applications
    • Information Systems and Management
    • Artificial Intelligence

    Cite this

    Adaptive recommendation model using meta-learning for population-based algorithms. / Chu, Xianghua; Cai, Fulin; Cui, Can; Hu, Mengqi; Li, Li; Qin, Quande.

    In: Information Sciences, Vol. 476, 01.02.2019, p. 192-210.

    Research output: Contribution to journalArticle

    Chu, Xianghua ; Cai, Fulin ; Cui, Can ; Hu, Mengqi ; Li, Li ; Qin, Quande. / Adaptive recommendation model using meta-learning for population-based algorithms. In: Information Sciences. 2019 ; Vol. 476. pp. 192-210.
    @article{fe24fbd566cd41c3b6a1cfa6270f09f9,
    title = "Adaptive recommendation model using meta-learning for population-based algorithms",
    abstract = "To efficiently solve complex optimization problems, numerous population-based meta-heuristics and extensions have been developed. However, the performances of the algorithms vary depending on the problems. In this research, we propose an Adaptive Recommendation Model (ARM) using meta-learning to identify appropriate problem-dependent population-based algorithm. In ARM, the algorithms are adaptively selected by mapping the problem characteristics to the algorithm performance. Since the meta-features extracted and meta-learner adopted would significantly affect the system performance, 18 meta-features including statistical, geometrical and landscape features are extracted to characterize optimization problem spaces. Both instance-based and model-based learners are investigated. Two performance metrics, Spearman's rank correlation coefficient and success rate are used to evaluate the accuracy of optimizer's ranking prediction and the precision of the best optimizer recommendation. The proposed ARM is compared against population-based algorithms with distinct search capabilities such as PSO variants, non-PSO population-based optimizers, hyper-heuristics and ensemble methods. Benchmark functions and real-world problems with various properties are adopted in the experiments. Experimental results reveal the extendibility and effectiveness of ARM on the diverse tested problems in terms of solution accuracy, ranking and success rate.",
    keywords = "Algorithm selection, Global optimization, Meta-learning, Population-based algorithm, Recommendation system",
    author = "Xianghua Chu and Fulin Cai and Can Cui and Mengqi Hu and Li Li and Quande Qin",
    year = "2019",
    month = "2",
    day = "1",
    doi = "10.1016/j.ins.2018.10.013",
    language = "English (US)",
    volume = "476",
    pages = "192--210",
    journal = "Information Sciences",
    issn = "0020-0255",
    publisher = "Elsevier Inc.",

    }

    TY - JOUR

    T1 - Adaptive recommendation model using meta-learning for population-based algorithms

    AU - Chu, Xianghua

    AU - Cai, Fulin

    AU - Cui, Can

    AU - Hu, Mengqi

    AU - Li, Li

    AU - Qin, Quande

    PY - 2019/2/1

    Y1 - 2019/2/1

    N2 - To efficiently solve complex optimization problems, numerous population-based meta-heuristics and extensions have been developed. However, the performances of the algorithms vary depending on the problems. In this research, we propose an Adaptive Recommendation Model (ARM) using meta-learning to identify appropriate problem-dependent population-based algorithm. In ARM, the algorithms are adaptively selected by mapping the problem characteristics to the algorithm performance. Since the meta-features extracted and meta-learner adopted would significantly affect the system performance, 18 meta-features including statistical, geometrical and landscape features are extracted to characterize optimization problem spaces. Both instance-based and model-based learners are investigated. Two performance metrics, Spearman's rank correlation coefficient and success rate are used to evaluate the accuracy of optimizer's ranking prediction and the precision of the best optimizer recommendation. The proposed ARM is compared against population-based algorithms with distinct search capabilities such as PSO variants, non-PSO population-based optimizers, hyper-heuristics and ensemble methods. Benchmark functions and real-world problems with various properties are adopted in the experiments. Experimental results reveal the extendibility and effectiveness of ARM on the diverse tested problems in terms of solution accuracy, ranking and success rate.

    AB - To efficiently solve complex optimization problems, numerous population-based meta-heuristics and extensions have been developed. However, the performances of the algorithms vary depending on the problems. In this research, we propose an Adaptive Recommendation Model (ARM) using meta-learning to identify appropriate problem-dependent population-based algorithm. In ARM, the algorithms are adaptively selected by mapping the problem characteristics to the algorithm performance. Since the meta-features extracted and meta-learner adopted would significantly affect the system performance, 18 meta-features including statistical, geometrical and landscape features are extracted to characterize optimization problem spaces. Both instance-based and model-based learners are investigated. Two performance metrics, Spearman's rank correlation coefficient and success rate are used to evaluate the accuracy of optimizer's ranking prediction and the precision of the best optimizer recommendation. The proposed ARM is compared against population-based algorithms with distinct search capabilities such as PSO variants, non-PSO population-based optimizers, hyper-heuristics and ensemble methods. Benchmark functions and real-world problems with various properties are adopted in the experiments. Experimental results reveal the extendibility and effectiveness of ARM on the diverse tested problems in terms of solution accuracy, ranking and success rate.

    KW - Algorithm selection

    KW - Global optimization

    KW - Meta-learning

    KW - Population-based algorithm

    KW - Recommendation system

    UR - http://www.scopus.com/inward/record.url?scp=85055089971&partnerID=8YFLogxK

    UR - http://www.scopus.com/inward/citedby.url?scp=85055089971&partnerID=8YFLogxK

    U2 - 10.1016/j.ins.2018.10.013

    DO - 10.1016/j.ins.2018.10.013

    M3 - Article

    VL - 476

    SP - 192

    EP - 210

    JO - Information Sciences

    T2 - Information Sciences

    JF - Information Sciences

    SN - 0020-0255

    ER -