Multi-objective adaptive differential evolution for SVM/SVR hyperparameters selection

Highlights：

• A novel metaheuristic has been proposed that combines an Adaptive Parameter control with a Mutant tournament in a Multi-Objective Differential Evolution (APMT-MODE).

• APMT-MODE is addressed for solving the Parameters Selection Problem (PSP) for SVM/SVRs.

• APMT-MODE uses the Criticism of Lexicographic Ordering (CLO) to adapt the Differential Evolution (DE) with a multi-objective approach to properly obtain solutions yielding a better tradeoff between high precision and a low number of support vectors.

• APMT-MODE yields hyperparameters, aiming to minimize the empirical risk and complexity, reducing the probability of underfitting and overfitting.

摘要

•A novel metaheuristic has been proposed that combines an Adaptive Parameter control with a Mutant tournament in a Multi-Objective Differential Evolution (APMT-MODE).•APMT-MODE is addressed for solving the Parameters Selection Problem (PSP) for SVM/SVRs.•APMT-MODE uses the Criticism of Lexicographic Ordering (CLO) to adapt the Differential Evolution (DE) with a multi-objective approach to properly obtain solutions yielding a better tradeoff between high precision and a low number of support vectors.•APMT-MODE yields hyperparameters, aiming to minimize the empirical risk and complexity, reducing the probability of underfitting and overfitting.

论文关键词：Support vector machines,Parameters selection problem,Multi-objective optimization,Differential evolution,Adaptive parameters strategy

论文评审过程：Received 28 October 2019, Revised 31 August 2020, Accepted 7 September 2020, Available online 11 September 2020, Version of Record 17 September 2020.