Navegando por Palavras-chave "Symbolic Regression"

Agora exibindo 1 - 1 de 1
  • Nenhuma Miniatura disponível
    Item
    Somente Metadadados
    Programação Kaizen Para Construção De Modelos Interpretáveis: Uma Abordagem Multiobjetivo Para Regressão Simbólica
    (Universidade Federal de São Paulo (UNIFESP), 2017-04-13) Alves, Artur Henrique Goncalves Coutinho [UNIFESP]; Melo, Vinicius Veloso De [UNIFESP]; Universidade Federal de São Paulo (UNIFESP)
    Regression problems are among some of the main current uses for machine learning and pattern recognition techniques, coming from the necessity of identifying relationships between behaviors and explanatory variables in economy, engineering, natural environment and numerous other areas. Traditional machine learning techniques such as artificial neural networks are extensively applied to real problems with good success rates, but present dificulties, such as parameter configuration, and shortcomings, such as the impossibility of interpreting the relationships found in the modeling of a specific system. Enters, then, the symbolic regression, a research subarea focused on methods for building mathematical equations for the correct modeling of different behaviors. This approach is not limited by incorrect structure choices, unlike linear regression, and allows for the analysis of the behavior modeled by the chosen mathematical elements. In this work, a recent automatic programming technique that can be used for symbolic regression is presented: Kaizen Programming. This technique applies continuous improvement concepts in a hyperheuristics structure, allowing for its use in various problems and with various auxiliary heuristics. Besides, it uses deterministic methods to evaluate and decide upon the ideas proposed by these techniques, lessening the negative impact a purely stochastic approach may bring to this kind of application. The implementation used here presents new modifications, specially the inclusion of a new objective: their complexity, defined by the nonlinearity of their mathematical elements. In real problems, it is expected that high quality models will be complex, but not too much, in order to avoid overfitting and to keep interpretability; therefore, the symbolic regression becomes a multiobjective problem, with conflicting objectives. This new version of Kaizen Programming was compared to the original one, to classical machine learning techniques and to another symbolic regression technique in well-known datasets constructed from real problems and in a time series - building autoregressive models for a predictive control to automatically drive a vehicle in a racing simulator. In general, the new technique presents lower predictive power when compared to its original counterpart and the other symbolic regression technique considered here, but offers solutions that are considerably simpler than the ones built by both.